One agent doing the work well is the default. Splitting is a tool for when it measurably helps: independent pieces, output too large for one context, or a tool that does one thing much better. Every split costs tokens, latency and a merge, and parallel agents make implicit decisions that can collide.
Decide first
- Do it directly when the change is small, when there is back and forth with the person, when every piece needs the same heavy context, or when latency matters. Delegating a five minute task costs more than doing it.
- Split when pieces are independent (no shared files or state, fixing one does not fix another), when research would flood the main context, or when a specialist tool clearly wins.
- Scale effort to the task: a single agent with a few calls for a fact, two to four workers for a comparison, more only for wide research or many independent files. A multi agent run uses many times the tokens of one agent.
The brief
Every brief is self contained; the worker has none of this conversation. It names:
- Goal and why, in two lines.
- Scope: the files or areas it owns, and what it must not touch.
- The project skill and the rules that apply, by path.
- Acceptance: an observable check (a test, a command output, a screenshot, an audit run) and what counts as done.
- Return: the exact shape of the report, short, with paths to artifacts instead of their content, ending in one status:
DONE,DONE_WITH_CONCERNS,NEEDS_CONTEXTorBLOCKED.
Constraints are stated in the negative when they matter ("do not raise timeouts", "do not change the public API").
Running workers
- Launch every independent worker at once, then wait. Waiting between launches serializes work that was meant to be parallel.
- Isolate writers. Two workers never edit the same checkout: use worktrees (
isolation: worktreein Claude Code,cursor-agent -w, Orca--worktree) or split by file ownership. - Pick the model on purpose. A fast, cheap model for mechanical or high volume work; the most capable one for design, judgment and final review. After the run, check which model actually ran.
- Long work is launched and checked apart. Never block a session waiting inside the command that does the work. A timeout or an empty wait is a checkpoint, not a failure; only positive proof that a process ended justifies stopping or retrying it.
- Keep a progress file for runs longer than one sitting, so a compacted context can resume instead of restarting.
Receiving work
- Do not trust the summary. Read the diff, open the artifact, run the acceptance check. "Done" from a worker is a claim, not evidence.
- Review in a fresh context: first against the brief (did it do what was asked, only that), then for quality. Flag only what affects correctness or the requirements.
- After merging parallel work, run the whole check again. Two correct pieces can be wrong together.
- On failure, resume before restarting. Retry the same worker once or twice with the missing context; after that, hand it to a fresh worker with a more capable model.
- The coordinator does not quietly fix a worker's output. It sends the correction back or does it openly and says so.
Who to delegate to
Use what is installed and allowed; skip the rest without stopping. Before calling an external tool, check it exists (--help or --version). Anything that spends money, acts on real accounts or opens windows on the person's machine is confirmed with them first.
| Need | First choice | Optional |
|---|---|---|
| Parallel research or code inside the session | Native sub agents: Claude Code (.claude/agents/*.md), Codex (.codex/agents/*.toml), Cursor (.cursor/agents/*.md) |
|
| Parallel workers, each in its own worktree, supervised from one place; task graphs, decision gates | Orca, the workspace this stack is built in. Check orca status; then load orca skills get orchestration for supervised runs, or orca skills get orca-cli to hand a task to an agent in its own worktree. Install the skills with orca skills install --skill orchestration --skill orca-cli. Without Orca, native sub agents with worktrees |
|
| Independent coding in its own worktree | Cursor CLI: cursor-agent -p --output-format json -w <name> "<brief>"; Cursor cloud agents through its API for work that ends in a PR |
|
| A second opinion or a different model on code | codex exec "<brief>" -C <dir> --json (GPT-6 Astra and GPT-5.x where available); grok -p "<brief>" --output-format json; gemini -p "<brief>" --output-format json |
|
| Images | OpenAI Images API (gpt-image-2.5-*); Higgsfield CLI (higgsfield generate create <model> --prompt ... --wait --json, check generate cost first) |
|
| Video | Higgsfield CLI | |
| Errands on real accounts (forms, returns, bookings) | Grok Bot. It has no API: write the task as one clear message for the person to give it. Never let it pay without the person's confirmation |
Flags change between releases; when a command fails, read its --help before guessing. External agents get the same brief, the same isolation and the same review as native ones.
Never
- Delegate to avoid reading the code. The coordinator understands the change it is splitting.
- Leave a worker undecided at the end: reuse it, keep it with a reason, or release it.
- Make the person a messenger between agents. If a handoff is needed, measure the risk; when there is none, do it and say what was touched.
- Report a parallel run as faster or better without the numbers to show it.